Blood Cell Classification via CNN and Hard Negative Mining

  • A blood cell classification solution for a famous medical device manufacturer in China.
  • The solution is based on CNN. Hard Negative Mining is adopted to achieve accuracy over 90%.
    project page  


Lenovo DeepNEX: A Distributed Multi-tenant Private Cloud for Deep Learning Development

  • Integrated deep learning toolkits e.g. Caffe, TensorFlow etc. into Docker container for use in the cloud.
  • Refactored deep learning projects for demonstration to our customers.
    project page  


Bone Age Analysis via Deep Regression Networks

  • A bone age analysis solution for Shanghai Children’s Hospital, China.
  • Developed two deep regression networks to auto-detect hand joints in X-ray images, and predict ages based on the ROIs of the joints.
    project page     source code (the annotation tool)  


CAPTCHA Cracking: A CNN Based OCR Module for Web Crawler.

  • Extended single-label classification to multi-label classification in CNN, and applied it into text recognition in a single image as CAPTCHA.
  • An end-to-end solution which avoids image segmentation for a single character.
  • Augmented synthesized data, to avoid overfitting when model applied to real data.
    project page     source code  


3D Camera: RGBD Image Algorithms for Dual-camera Mobile Phone, Lenovo VIBE S1.

  1. Image Refocus
    • Developed an image refocus algorithm which utilizes the depth information to blur the image to simulate an effect of “large aperture, shallow depth of field”.
    • Developed efficient algorithms and applied parallel computing (OpenCL) to achieve real-time interaction.
  2. Image Matting
    • Co-developed an automatic selfie image matting algorithm. Over-segmentation, region-wise matting and parallel processing are involved to save memory and speed up the program.
    project page     video demo  


AR Camera: An Augmented Reality Prototype for Mobile Devices of Lenovo.

  • Prototyped an application with an AR effect for QR code or a dish of food, to improve user experience.
  • Developed detection, tracking and stereo algorithms to obtain a real-time and smooth effect.
    video demo  


Image Matting Application.

  • Developed an interactive application to cut foreground from an image based on Robust Image Matting algorithm.
    project page     source code